Repository navigation
Expand file tree
/
Copy pathsharp_positions.py
More file actions
447 lines (349 loc) · 15.9 KB
/
Copy pathsharp_positions.py
File metadata and controls
447 lines (349 loc) · 15.9 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
# Copyright 2018 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Used as a starting point for our generating hard position collection.
training_curve.py is a supervised approach, assuming the pro moves are correct
Performance is then measured against the played move and eventual outcome.
sharp_positions is unsupervised, it tries to determine what the correct outcome
and move is based on clustering of the strongest rated algorithms.
Step 1. Create collection.csv of sgf and move number
export BOARD_SIZE=19
SGF_DIR=data/sgf
MODEL_DIR=models/
python3 sharp_positions.py subsample --num_positions 10 --sgf_dir $SGF_DIR
Step 2. Create a directory of those sgfs truncated at the specified move number
# from https://github.com/sethtroisi/go-scripts/
./truncate.sh
# Remove any bad SGFS.
cat badfiles.txt | tqdm | xargs rm
# Remove original SGFs
cat ../minigo/collection.csv | cut -f1 -d, | sort -u | tqdm | xargs rm
# Rerun truncate successfully this time
./truncate.sh
Step 3. Get value & policy for all models for all positions
python3 sharp_positions.py evaluate --sgf_dir <problem-collections2> --model_dir $MODEL_DIR
Step 4. Fit a model and minimize a set of positions to predict strength
python3 sharp_positions.py minimize \
--model_dir $MODEL_DIR --sgf_dir data/s \
--rating_json ratings.json --results results.csv
"""
import sys
sys.path.insert(0, '.')
import itertools
import json
import multiprocessing
import os
import random
from collections import defaultdict, Counter
import fire
import numpy as np
import tensorflow as tf
from absl import flags
from sklearn import svm
from tqdm import tqdm
import oneoff_utils
# Filter tensorflow info
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '1'
flags.DEFINE_integer("num_positions", 1, "How many positions from each game.")
flags.DEFINE_integer("top_n", 3, "Policy moves to record per position.")
flags.DEFINE_integer("min_idx", 100, "Min model number to include.")
flags.DEFINE_integer("batch_size", 64, "Eval batch size.")
# Inputs
flags.DEFINE_string("sgf_dir", "data/s/", "input collection of SGFs")
flags.DEFINE_string("model_dir", "models/", "Folder of Minigo models")
flags.DEFINE_string("rating_json", "ratings.json", "Ratings of models")
# Outputs
flags.DEFINE_string("collection", "collection.csv", "subsampled csv file")
flags.DEFINE_string("results", "results.csv", "Evaluate results file")
flags.DEFINE_string("SVM_json", "SVM_data.json", "SVM data about positions")
FLAGS = flags.FLAGS
def grouper(n, iterable):
"""Itertools recipe
>>> list(grouper(3, iter('ABCDEFG')))
[['A', 'B', 'C'], ['D', 'E', 'F'], ['G']]
"""
return (iterable[i:i + n] for i in range(0, len(iterable), n))
def get_final_positions():
sgf_files = sorted(oneoff_utils.find_and_filter_sgf_files(FLAGS.sgf_dir))
with multiprocessing.Pool() as pool:
pos = list(pool.map(oneoff_utils.final_position_sgf, tqdm(sgf_files)))
assert len(pos) > 0, "BOARD_SIZE != 19?"
return sgf_files, pos
def get_model_idx(model_name):
number = os.path.basename(model_name).split('-')[0]
assert len(number) == 6 and 0 <= int(number) <= 1000, model_name
return int(number)
def subsample():
"""Sample num_positions postions from each game in sgf_dir
Usage:
python3 sharp_positions.py subsample --num_positions 10 --sgf_dir data/s
NOTE(sethtroisi): see link for a script to truncate SGFs at move number
https://github.com/sethtroisi/go-scripts
"""
sgf_files = oneoff_utils.find_and_filter_sgf_files(
FLAGS.sgf_dir, None, None)
with open(FLAG.collection, 'w') as collection:
fails = 0
for path in tqdm(sorted(sgf_files)):
try:
positions, moves, results = oneoff_utils.parse_sgf(path)
except KeyboardInterrupt:
raise
except Exception as e:
fails += 1
print("Fail {}, while parsing {}: {}".format(fails, path, e))
continue
moves = len(positions)
indexes = random.sample(range(10, moves), FLAGS.num_positions)
for index in sorted(indexes):
collection.write('{}, {}\n'.format(path, index))
def evaluate():
"""Get Policy and Value for each network, for each position
Usage:
python3 sharp_positions.py evaluate --sgf_dir data/s --model_dir models/
"""
def short_str(v):
if isinstance(v, float):
return "{.3f}".format(v)
return str(v)
# Load positons
sgf_names, all_positions = get_final_positions()
# Run and save some data about each position
# Save to csv because that's easy
model_paths = oneoff_utils.get_model_paths(FLAGS.model_dir)
num_models = len(model_paths)
print("Evaluating {} models: {} to {}".format(
num_models, model_paths[0], model_paths[-1]))
print()
with open(FLAGS.results, "w") as results:
results.write(",".join(sgf_names) + "\n")
player = None
for idx in tqdm(range(FLAGS.min_idx, num_models, 1), desc="model"):
model = model_paths[idx]
if player and idx % 50 == 0:
player.network.sess.close()
tf.reset_default_graph()
player = None
if player:
oneoff_utils.restore_params(model, player)
else:
player = oneoff_utils.load_player(model)
row = [model]
for positions in grouper(FLAGS.batch_size, all_positions):
probs, values = player.network.run_many(positions)
# NOTE(sethtroisi): For now we store the top n moves to shrink
# the size of the recorded data.
top_n = FLAGS.top_n
top_policy_move = np.fliplr(np.argsort(probs))[:,:top_n]
top_policy_value = np.fliplr(np.sort(probs))[:,:top_n]
# One position at a time
for v, m, p in zip(values, top_policy_move, top_policy_value):
row.append(v)
row.extend(itertools.chain.from_iterable(zip(m, p)))
if len(positions) > 10:
average_seen = top_policy_value.sum() / len(positions)
if average_seen < 0.3:
print("\t", average_seen, top_policy_value.sum(axis=-1))
results.write(",".join(map(short_str, row)) + "\n")
def minimize():
"""Find a subset of problems that maximal explains rating.
Usage:
python3 sharp_positions.py minimize \
--model_dir models --sgf_dir data/s
--rating_json ratings.json --results results.csv
"""
########################### HYPER PARAMETERS ###############################
# Stop when r2 is this much worse than full set of positions
r2_stopping_percent = 0.96
# for this many iterations
stopping_iterations = 5
# Limit SVM to a smaller number of positions to speed up code.
max_positions_fit = 300
# Filter any position that "contributes" less than this percent of max.
filter_contribution_percent = 0.3
# Never filter more than this many positions in one iterations
filter_limit = 25
########################### HYPER PARAMETERS ###############################
# Load positons
model_paths = oneoff_utils.get_model_paths(FLAGS.model_dir)
num_models = len(model_paths)
assert num_models > 0, FLAGS.model_dir
# Load model ratings
# wget https://cloudygo.com/v12-19x19/json/ratings.json
ratings = json.load(open(FLAGS.rating_json))
raw_ratings = {int(r[0]): float(r[1]) for r in ratings}
model_ratings = []
for model in model_paths:
model_idx = get_model_idx(model)
if model_idx < FLAGS.min_idx:
continue
model_ratings.append(raw_ratings[model_idx])
model_ratings = np.array(model_ratings)
assert 0 < len(model_ratings) <= num_models, len(model_ratings)
num_models = len(model_ratings)
sgf_names, all_positions = get_final_positions()
# Trim off common path prefix.
common_path = os.path.commonpath(sgf_names)
sgf_names = [name[len(common_path) + 1:] for name in sgf_names]
print("Considering {} positions, {} models".format(
len(all_positions), num_models))
print()
# Load model data
top_n = FLAGS.top_n
positions = defaultdict(list)
with open(FLAGS.results) as results:
headers = results.readline().strip()
assert headers.count(",") + 1 == len(sgf_names)
# Row is <model_name> + positions x [value, top_n x [move, move_policy]]
for row in tqdm(results.readlines(), desc="result line"):
data = row.split(",")
model_idx = get_model_idx(data.pop(0))
if model_idx < FLAGS.min_idx:
continue
data_per = 1 + top_n * 2
assert len(data) % data_per == 0, len(data)
for position, position_data in enumerate(grouper(data_per, data)):
value = float(position_data.pop(0))
moves = list(map(int, position_data[0::2]))
move_policy = list(map(float, position_data[1::2]))
positions[position].append([value, moves, move_policy])
def one_hot(n, i):
one_hot = [0] * n
if 0 <= i < n:
one_hot[i] += 1
return one_hot
# NOTE: top_n isn't the same semantic value here and can be increased.
one_hot_moves = top_n
num_features = 1 + 5 + (one_hot_moves + 1)
# Features by position
features = []
pos_top_moves = []
for position, data in tqdm(positions.items(), desc="featurize"):
assert len(data) == num_models, len(data)
top_moves = Counter([d[1][0] for d in data])
top_n_moves = [m for m, c in top_moves.most_common(one_hot_moves)]
if len(top_n_moves) < one_hot_moves:
top_n_moves.extend([-1] * (one_hot_moves - len(top_n_moves)))
assert len(top_n_moves) == one_hot_moves, "pad with dummy moves"
pos_top_moves.append(top_n_moves)
# Eventaully we want
# [model 1 position 1 features, m1 p2 features, m1 p3 features, ... ]
# [model 2 position 1 features, m2 p2 features, m2 p3 features, ... ]
# [model 3 position 1 features, m3 p2 features, m3 p3 features, ... ]
# ...
# [model m position 1 features, mm p2 features, mm p3 features, ... ]
# We'll do position selection by joining [model x position_feature]
feature_columns = []
for model, (v, m, mv) in enumerate(data):
# Featurization (for each positions):
# * Value (-1 to 1), Bucketed value
# * Cluster all model by top_n moves (X,Y,Z or other)?
# * value of that move for model
# * policy value of top move
model_features = []
model_features.append(2 * v - 1)
# NOTE(sethtroisi): Consider bucketize value by value percentiles.
value_bucket = np.searchsorted((0.2, 0.4, 0.6, 0.8), v)
model_features.extend(one_hot(5, value_bucket))
# Policy weight for most common X moves (among all models).
policy_weights = [0] * (one_hot_moves + 1)
for move, policy_value in zip(m, mv):
if move in top_n_moves:
policy_weights[top_n_moves.index(move)] = policy_value
else:
policy_weights[-1] += policy_value
model_features.extend(policy_weights)
assert len(model_features) == num_features
feature_columns.append(model_features)
features.append(feature_columns)
features = np.array(features)
print("Feature shape", features.shape)
print()
# Split the models to test / train
train_size = int(num_models * 0.9)
train_models = sorted(np.random.permutation(num_models)[:train_size])
test_models = sorted(set(range(num_models)) - set(train_models))
assert set(train_models + test_models) == set(range(num_models))
features_train = features[:, train_models, :]
features_test = features[:, test_models, :]
labels_train = model_ratings[train_models]
labels_test = model_ratings[test_models]
# Choose some set of positions and see how well they explain ratings
positions_to_use = set(positions.keys())
linearSVM = svm.LinearSVR()
best_test_r2 = 0
below_threshold = 0
for iteration in itertools.count(1):
iter_positions = np.random.permutation(list(positions_to_use))
iter_positions = sorted(iter_positions[:max_positions_fit])
# Take this set of positions and build X
X = np.concatenate(features_train[iter_positions], axis=1)
Xtest = np.concatenate(features_test[iter_positions], axis=1)
assert X.shape == (train_size, num_features * len(iter_positions))
linearSVM.fit(X, labels_train)
score_train = linearSVM.score(X, labels_train)
score_test = linearSVM.score(Xtest, labels_test)
print("iter {}, {}/{} included, R^2: {:.4f} train, {:.3f} test".format(
iteration, len(iter_positions), len(positions_to_use),
score_train, score_test))
# Determine the most and least useful position:
# TODO(amj,brilee): Validate this math.
assert len(linearSVM.coef_) == num_features * len(iter_positions)
# The intercepts tell us how much this contributes to overall rating
# but coef tell us how much different answers differentiate rating.
coef_groups = list(grouper(num_features, linearSVM.coef_))
position_coefs = [abs(sum(c)) for c in coef_groups]
pos_value_idx = np.argsort(position_coefs)
max_pos = pos_value_idx[-1]
most_value = position_coefs[max_pos]
print("\tMost value {} => {:.1f} {}".format(
max_pos, most_value, sgf_names[iter_positions[max_pos]]))
# Drop any positions that aren't very useful
for dropped, pos_idx in enumerate(pos_value_idx[:filter_limit], 1):
contribution = position_coefs[pos_idx]
positions_to_use.remove(iter_positions[pos_idx])
print("\t\tdropping({}): {:.1f} {}".format(
dropped, contribution, sgf_names[iter_positions[pos_idx]]))
if contribution > filter_contribution_percent * most_value:
break
print()
best_test_r2 = max(best_test_r2, score_test)
if score_test > r2_stopping_percent * best_test_r2:
below_threshold = 0
else:
below_threshold += 1
if below_threshold == stopping_iterations:
print("{}% decrease in R^2, stopping".format(
100 - int(100 * r2_stopping_percent)))
break
# Write down the differentiating positions and their answers.
svm_data = []
for position_idx in list(reversed(pos_value_idx)):
coefs = coef_groups[position_idx]
# Global position index.
position = iter_positions[position_idx]
sgf_name = sgf_names[position]
top_moves = pos_top_moves[position]
svm_data.append([sgf_name, [top_moves, coefs.tolist()]])
with open(FLAGS.SVM_json, "w") as svm_json:
json.dump(svm_data, svm_json)
print("Dumped data about {} positions to {}".format(
len(svm_data), FLAGS.SVM_json))
if __name__ == "__main__":
remaining_argv = flags.FLAGS(sys.argv, known_only=True)
fire.Fire({
'subsample': subsample,
'evaluate': evaluate,
'minimize': minimize,
}, remaining_argv[1:])